Combining Spatial and Social Awareness in D2D Opportunistic Routing

IEEE COMMUNICATIONS MAGAZINE 1 Combining Spatial and Social Awareness in D2D Opportunistic Routing Ivan O. Nunes, Clayson Celes, Igor Nunes, Pedro O. S. Vaz de Melo, Antonio A. F. Loureiro Department of Computer Science Federal University of Minas Gerais Brazil Abstract Social-aware algorithms have remarked themselves as the most successful strategies for cost-effective content delivery in mobile opportunistic networks. However, these strategies do not consider the importance of users geographical locations nor the spatial properties of human mobility. Given this fact, in this work we propose to combine spatial and social properties to improve the cost-effectiveness of content delivery in opportunistic Deviceto-Device (D2D) networks. We leverage and describe four spatial and social properties and characterize them in a real-world dataset. As a proof of concept of employing those properties, we propose, a forwarding algorithm that combines social awareness, points of interest within a region, and users individual mobility patterns to cost-effectively deliver messages in opportunistic D2D networks. We compare to the state-of-the-art social-aware algorithm, namely, and with a modified version of that incorporates static relay nodes. Our experiments, conducted using both real-world (NCCU) and state-of-art synthetic (SWIM) traces, confirm that the combination of social and spatial awareness can increase the delivery performance. Keywords Opportunistic Routing, Device-to-Device Communication, Social Networking, Mobility. I. INTRODUCTION Opportunistic mobile networks have attracted a lot of attention and several algorithms have been proposed to enable cost-effective and timely delivery of data. In such networks, a given content is forwarded device-by-device, from the source to the destination, using the intermittently connected structure of mobile networks. More recently, there was an increasing use of high data rate applications in cellular networks that make use of heavy multimedia content such as videos, music, games, and social media. Opportunistic forwarding has been proposed to facilitate the high data rate transmissions among nearby users in these scenarios, offering the possibility of offloading the traffic demands of a base station. This type of communication might be used to deliver contents such as video advertisements and noncritical updates of applications. In such cases, timely delivery is not essential. This is known as opportunistic Device-to-Device (D2D) communication offloading in cellular networks [1], [2]. The goal of forwarding algorithms for opportunistic D2D cellular networks is to deliver as many messages as possible maintaining a low network overhead. Considering these metrics, [3], the most successful strategy for opportunistic cost-effective forwarding, relies on information about social communities and nodes centrality, which can be approximated by the node popularity within the mobile network. However, and previous social-aware strategies do not consider any geographic feature of the scenario in question, such as its Points of Interest (PoI), nor the users mobility patterns. The recently released NCCU trace [4] brings a remarkable opportunity to investigate this open issue, since it is the first available real-world dataset to monitor, not only users proximity contacts, but also their geo-locations. In the literature, there are other real traces, but none of them presents all these properties. In other words, based on the characterization of the NCCU trace we can obtain insights to design real-world protocols that take advantage and consider human mobility. With that in mind, in this work we propose to combine spatial features and social awareness, recorded in the NCCU trace, with the goal of improving the cost-effectiveness of opportunistic forwarding. We describe two spatial and two social features and characterize them in the NCCU trace. As a proof of concept, we use such properties to design (Social-Aware, Mobility, and PoI Routing) that, to the best of our knowledge, is the first opportunistic routing strategy to combine mobility, PoIs, and social-awareness to provide costeffective content delivery in intermittent networks. We evaluate using two traces: NCCU and the synthetic mobility model SWIM [5]. The latter allows the performance evaluation of a large network. works by forwarding messages to nodes of higher mobility, until the message reaches a static relay point. Static relay points are strategically deployed at the most popular PoIs and forward their received content to nodes that belong to the social community whose destination node is also a member. Within such community, the message is forwarded to the most popular nodes until it reaches the destination node. Our experiments show that, by exploring the combination of spatial and social features, improves delivery ratio, reduces the network overhead, and enables faster delivery of messages. These results reinforce our assumption that a better understanding of real mobility traces can provide valuable insights in the design of D2D opportunistic routing. II. BACKGROUND AND RELATED WORK D2D refers to the direct transmission of content between devices without the need of sending all data through a base

IEEE COMMUNICATIONS MAGAZINE 2 station, as in traditional cellular networks [1]. D2D communication can be classified in two basic types: 1-hop transmission, in which a message goes directly from the source to the destination, if they are close enough to each other; and multihop transmission, where the message must be opportunistically routed, device-by-device, from the source to the destination. The latter solution is more complex since it depends on the intermittent communication structure of a mobile network, being suitable for communications that tolerate large delivery times. This concept was initially introduced in the context of opportunistic networks, but can naturally be applied to D2D networks, acting as a bandwidth offload mechanism for the download demands of the base stations [1]. At first, oblivious solutions (in which no context information is considered) were proposed for such networks. Examples of forwarding strategies in this category are Flooding (or epidemic propagation) and Spray and Wait [6]. However, these strategies are not practical since the number of content retransmissions grow extremely fast as the number of nodes in the network increases. As a consequence, several opportunistic communication solutions were proposed to achieve costeffective delivery [1], [3], [7], [8], [9], [10]. In this case, the goal is to achieve the highest possible delivery ratio with the lowest possible network overhead. Delivery ratio is measured as the percentage of opportunistically routed messages successfully delivered to their destinations. These messages are those that the base station will not need to deliver itself, thus using less bandwidth. Network overhead is measured by the average number of times content needs to be D2D-transmitted for the message to get to its destination. A high number of transmissions may negatively impact the users experience by, for example, increasing the devices energy expenditure. Considering cost-effective solutions, the most successful approaches for opportunistic forwarding are the probabilistic and social-aware strategies. The use of a probabilistic approach was firstly introduced with the PROPHET algorithm [9]. Its main idea is to assign a higher importance to pairwise node contacts that happened more recently, in an attempt to predict future pair contacts. This algorithm achieved great success, but years later was outperformed by [3], which uses the community structure of mobile social networks of encounters, together with the knowledge about the nodes popularity, to define its forwarding policy. proposed a costeffective forwarding algorithm in terms of successful message delivery and low network overhead. More recently, a trend in the design of protocols consists in combining features to improve the decision-making process of routing. For example, SCORP [8] and Oi [7] combine users social ties with their interests for social-aware opportunistic routing. In [10], we have proposed GROUPS-Net, an algorithm that outperformed by combining social awareness with a probabilistic approach using group meetings as a measure of social context. An extensive review of opportunistic routing strategies is provided in [11]. Although aforementioned solutions have incrementally reduced the overhead and improved the delivery ratio, they do not combine spatial and social features to perform D2D routing. As mentioned earlier, human mobility presents spatial and social features that once characterized can be used to design new opportunistic forwarding algorithms. In this work, we discuss these features and their importance, and propose a novel protocol, namely, that makes use of such features. To the best of our knowledge, is the first forwarding algorithm to consider the combination of both spatial and social features. In Table I we show a summary of the solutions we have discussed. Property Algorithm Probabilistic Social PoI Individual Model Mobility Flooding Spray and Wait [6] [9] SCORP [8] [3] GROUPS-Net [10] TABLE I. PROPERTIES CONSIDERED IN THE MAIN OPPORTUNISTIC FORWARDING PROTOCOLS III. SPATIAL AND SOCIAL FEATURES The main contribution of this work is to show that it is worth combining information extracted from the user s geo-locations with social context to improve the state-of-the-art forwarding strategies. In this section, we individually explain four spatial and social features present in human mobility and revealed in a real trace we characterize and explore. A. The NCCU Trace The usual goal of using mobility traces in the evaluation of communication protocols is to capture the performance of such protocols in real-world scenarios, allowing more reliable analysis. Synthetic mobility models cannot capture all properties of human mobility and, therefore, may lead to biased results. In the case of opportunistic networking, this is especially harmful, because most of the existing mobility models do not capture the impact of human social bonds in the mobility. In this work, we combine mobility and social properties, based on the characteristics captured from a real-world mobility trace. The NCCU trace registered 115 users moving on a campus throughout a 15-day period. An Android application collected their GPS data, application usage, Wi-Fi access points, and proximity information captured via Bluetooth. These data were registered once every 10 minutes and contacts were detected when two users were less than 10 meters apart. To the best our knowledge, NCCU is the first public dataset that recorded pairwise contact information and users geo-locations within a dense mobility scenario. The NCCU trace provides a new opportunity to combine multiple features in several types of mobility studies. In this work, we are especially interested in providing insights on how to use spatial and social features to design an opportunistic forwarding algorithm for cost-effective routing in D2D networks. With that goal, we extract NCCU users mobility, popularity, PoIs, and social bonds. Below, we discuss each one of these properties and how they can be explored.

IEEE COMMUNICATIONS MAGAZINE 3 (a) Popularity of network users (b) Nine social communities detected from the (c) Trajectories of users with different radius of NCCU contact graph using the clique percolation gyrations method (each color represents a different community) (d) Users radius of gyration (P.D.F.) (e) Heat map of the NCCU mobility trace depicting the most popular places (f) Strategic deployment of four relay points with reach radius of 30m (red circles) Fig. 1. Mobility, PoIs and social features computed from the NCCU trace B. Social Awareness as Popularity Currently, the most successful approaches for opportunistic forwarding are the social-aware strategies [1]. These solutions aim to identify the most popular nodes in order to use them when forwarding messages. In summary, most popular nodes (persons) are those who meet others more often. The C- Window metric [3] measures nodes popularity as the average number of different nodes encountered throughout time windows of fixed length, e.g., 24 hours. C-Window presented high correlation (up to 0.95 correlation) with the node s betweenness centrality in the mobile network [3]. Nodes betweenness centrality in opportunistic networks is defined as the number of times a given node belongs to the shortest path (the one with fewer re-transmissions) between two nodes. Thus, it makes perfect sense to use such information to decide whether to forward or not a message to an encountered node, i.e., a message should be preferentially forwarded to more popular nodes. In human society, people have different levels of popularity. Hence, the main idea here is to use such heterogeneity to design more efficient forwarding schemes. For instance, Figure 1(a) presents the NCCU trace users popularity measured with the C-Window technique. It is possible to notice that few nodes have very high popularity, which means that a message forwarded to these strategically selected nodes have a high probability of reaching a given destination rapidly and with lower network overhead. C. Social Communities A second interesting approach to consider in the design of social-aware applications is the use of social communities. Communities in graphs are defined as groups of more densely interconnected nodes. By using social graphs, in which nodes are the network users and edges weights the number of meetings between pairs of nodes, it is possible to detect social communities. Among the algorithms for community detection in graphs, the Clique Percolation Method (CPM) [12] has remarked itself as one of the most effective methods, if set with

IEEE COMMUNICATIONS MAGAZINE 4 the correct parameters [13]. Figure 1(b) depicts nine social communities detected from the NCCU social contact graph using the CPM. The use of social communities in opportunistic forwarding is interesting because a given message has a much higher probability to be delivered to the destination once it gets to any member of a community where the destination belongs to. As we observe in Figure 1(b), nodes within communities are more densely interconnected, as intuitively expected, and according to the definition of communities in graphs. D. Users Individual Mobility Several studies about human mobility have shown that people have different mobility patterns, which are mostly influenced by their daily routines. In this regard, a more detailed understanding of people s mobility can provide significant insights toward the design of opportunistic routing solutions. There are studies in the literature that propose a single parameter to characterize the individual mobility of people. Radius of gyration [14] quantifies the dynamic mobility of a person in relation to the center of mass of his/her movements. Radius of gyration (r g ) of a given person is computed as in Equation 1, where n is the total number of recorded positions for a given user, p i represents the ith position recorded for the user, and p center is the center of mass of the user s recorded displacements, obtained as in Equation 2. r g = 1 n (p i p center ) 2, (1) n i=1 p center = 1 n n p i. (2) i=1 Figure 1(d) depicts the Probability Distribution Function (P.D.F.) of users radius of gyrations in the NCCU trace. We can see that most users have small radius of gyration (between 100 and 250 meters) and a few of them present a much higher mobility (above 400 meters). The radius of gyration is associated with the user s displacement, and, thus, it makes sense to forward messages to nodes with higher radius of gyration to increase the geographical coverage of a message in the network without needing to transmit copies of the messages to many nodes. Figure 1(c) presents the trajectories of three different users of the NCCU trace. User 50 exemplifies a person with high radius of gyration, covering a considerable region. Conversely, User 43 has a low radius of gyration, covering a limited region. E. Points of Interest Despite people presenting different behaviors, their mobility has some spatial-temporal intersections. For instance, it is common to see very distinct people (e.g., students, executives and retirees) being attracted constantly to the same geographical location (e.g., train station). Such locations, called Points of Interest (PoI), may be used for the deployment of static relay nodes. Each static relay node consists of a low-computational power device with communication and storage capabilities, similar to access points, but with no need for Internet connection. Its role in the network is to store and forward messages according to a given forwarding policy. Due to these characteristics, they have low financial cost. Figure 1(e) shows the heat map of the frequency of visitations in the NCCU campus. Considering this frequency, Figure 1(f) depicts the strategic deployment of four relay points with communication radius of 30 m (red circles). As we can observe, they were deployed at the most visited PoIs in the NCCU campus. As motivation, in Figure 2, we evaluate the performance of the flooding forwarding strategy when different numbers of relay points are added. The performance is measured by the delivery ratio (Figure 2(a)) and delivery time (Figure 2(b)). In the flooding forwarding strategy, every time a node holding a message meets another node that does not have it, the message is propagated. Flooding represents the upper bound for the delivery ratio, but drastically increases the network overhead. Nevertheless, notice that even in this simple strategy the deployment of static relay points presents benefits. Below, we discuss better ways to use static relay points to improve the delivery ratio without increasing the network overhead. It is important to mention that previous evaluations of the performance of static relay points, in opportunistic networks, only used synthetic mobility models [15]. However, as already discussed, those models do not capture human sociability, and, thus, the work in [15] did not consider the combination of PoIs with social awareness. IV. COMBINING FEATURES TOWARD COST-EFFECTIVE FORWARDING In the following, we present, which uses social communities and social popularity metrics as they were introduced in the original scheme [3], and adds the individual mobility and PoI properties to them in a clever way, helping to achieve high delivery ratio and reducing the network overhead. Since we want to quantify the improvement of adding these properties, we first present a brief overview of the algorithm. A. Algorithm [3] identifies social communities by looking at densely interconnected nodes in the aggregated contact graph, obtained from a given trace, using CPM [12]. Each node in the network must belong to at least one community. Nodes that do not belong to any community are assigned to a pseudocommunity of one node. This is necessary for the forwarding algorithm operation. Moreover, each node gets a measure of its global popularity in the network (GlobalRank) and a local measurement of popularity, which is valid within that node s community (LocalRank). Using these parameters, the forwarding strategy works as follows: At each encounter, a given node transmits its content if the encountered node has a higher GlobalRank or if the encountered node belongs to a community of which the final destination is a member.

IEEE COMMUNICATIONS MAGAZINE 5 (a) Delivery ratio (b) Delivery time distributions as boxplots Fig. 2. Effect of introducing relay points in the flooding forwarding Once the message is inside the final destination s community, the forwarding process occurs if the LocalRank of the encountered node is higher than the LocalRank of the node that has the message. This procedure goes on until the message reaches the final destination. Both the GlobalRank and the LocalRank are calculated using the C-Window technique that better approximates the node centrality within the mobile network. B. combines all social and spatial features that we have discussed earlier. To explore information about PoIs, static relay points are deployed at the most frequently visited places in the region of interest. The addition of the relay points by itself could be harmful, since it would drastically increase the number of message copies in the network, generating more network overhead. Because of that, combines PoIs and communities awareness to define the relay points forwarding policy. Instead of forwarding a message to any node that gets inside its transmission radius, a relay point only forwards messages to nodes that belong to the destination community, i.e., the community whose destination node is part of. Since relay points are placed in most popular areas (SAM- PLER s PoI feature), they will probably be in contact range with many different network nodes. Thus, the idea is to have them assuming the role of extremely popular nodes that accelerate the message forwarding to the destination community. As relay points have higher popularity, every mobile node always forwards a message when it meets a relay point. An obvious shortcoming of relay points is that they are static. To address this issue, adds Individual Mobility awareness to the forwarding scheme. Instead of setting the nodes GlobalRank as the nodes popularity, uses the nodes radius of gyration, which measures their average displacements ( s mobility feature). Thus, a message is forwarded to nodes that have higher mobility until it gets to a relay point. Once the message is inside the destination community, the forwarding policy works as in, transmitting the message to nodes with higher popularity within that community until the message gets to the destination ( s social features). In summary, works as follows: 1) GlobalRank is set to the nodes radius of gyration (mobility awareness); 2) LocalRank is set to the nodes popularity (C-Window metric) within the destination community (social awareness); 3) The message is always forwarded if the encountered node is the destination; 4) Every time a mobile node encounters a static relay point that does not have the message yet, the message is forwarded (PoI awareness); 5) Upon encountering a node that belongs to the destination community, the relay point forwards the message (social awareness); 6) Outside the destination community, the message is forwarded if the GlobalRank of the encountered node is higher than the GlobalRank of the node that has the message (mobility awareness). This step is performed at most once by each node; 7) Inside the destination community, the message is forwarded if the LocalRank of the encountered node is higher than the LocalRank of the node that has the message (social awareness). s key ideas are quite intuitive. Nodes with higher mobility help messages to get to relay points through a mobility-based forwarding strategy (Steps 1 and 6). Since relay points are placed in very popular areas, they meet many nodes, helping the message to get inside the destination s community faster and with less overhead (Steps 4 and 5). Finally, a popularity-based forwarding scheme is used within the destination community to deliver a message to its destination node (Steps 2 and 7). Figure 3 illustrates these principles.

IEEE COMMUNICATIONS MAGAZINE 6 purpose, in addition to, we propose Relay-Bubble, which only accounts for PoIs. In summary, we compute each of the aforementioned metrics for the following algorithms: Original : it does not use relay points (PoI awareness) nor the node s radius of gyration (mobility awareness). Relay-Bubble: variant of we proposed with relay points (PoI awareness) to get a message inside the destination community, and uses both GlobalRank and LocalRank as the node s popularity. : uses relay points (PoI awareness) to get a message inside the destination community, the radius of gyration as the GlobalRank (mobility awareness) and the LocalRank as the node s popularity (C-Window metric). : for completeness we also consider in our analysis, since it is the most cost-effective forwarding algorithm that does not use social information. In addition to NCCU, we also conducted the same experiment in the SWIM synthetic trace [5], with the goal of evaluating the performance of the algorithms in a larger scale network. SWIM is a state-of-the-art trace containing 700 nodes and capable of simulating mobility and social communities among its nodes. Fig. 3. Principles of the algorithm V. PERFORMANCE ANALYSIS A. Experimental methodology With the goal of comparatively evaluating and, we used the following traditional metrics: Delivery ratio: evaluates the percentage of successfully delivered messages for different values of Time To Live (TTL). Number of transmissions: measures the network overhead, i.e., the number of device-to-device transmissions that each algorithm performs for different TTLs. Delivery Time: measures the average time a message takes to get from the source node to its destination node. The algorithms were tested for different numbers of relay points. For each test campaign, we considered every possible (origin, destination) pair 1 for each algorithm and for each number of relay points. We have also evaluated separately the contribution of the relay points and the combination of relay points with the use of the radius of gyration in the GlobalRank parameter. With that 1 Since all possible (origin, destination) pairs were emulated, the experiments results are deterministic. Therefore, confidence intervals are not presented. B. Comparative Evaluation The algorithms were evaluated with several different numbers of relay points covering the k most popular PoIs. The results for k = 5, 10 are presented in Figure 4 for the NCCU trace and in Figure 5 for the SWIM trace. The first interesting observation is that s performance was significantly lower than the performance of the other three algorithms. In all considered scenarios, had lower delivery ratio, higher network overhead, and high delay to deliver messages. The only exception is Relay Bubble s overhead, the highest in the SWIM dataset, discussed later on. In both traces, takes more time to start delivering the first messages and when that happens the network overhead also starts to increase rapidly. This result confirms that the use of social awareness gives a drastic advantage, allowing protocols to perform much more efficiently. Considering the three social-aware strategies, the experiments in the NCCU trace show that Relay-Bubble has a performance similar to, with respect to the delivery ratio and transmission delay. Both of them present faster delivery when compared to (Figures 4(c) and 4(f)). Relay- Bubble also presents slightly faster delivery with 5 relay points when compared to. However, the most significant contrast between Relay-Bubble and is observed in the network overhead (Figures 4(b) and 4(e)). When comparing the number of transmissions that each algorithm performs over time, we notice that, for all considered numbers of relay points, Relay-Bubble is the most expensive strategy. On the other hand, due to the mobility awareness incorporated into, it achieves the fast delivery and high delivery ratio presented by Relay-Bubble with much lower network overhead.

IEEE COMMUNICATIONS MAGAZINE 7 Delivery Ratio (%) 0 20 40 60 80 100 # of Transmissions/Message 0 20 40 60 80 Avg. Delivery time(h) 0 50 100 150 Message TTL (a) Delivery 5 Relay Points (b) Transmissions 5 Relay Points (c) Delay 5 Relay Points Delivery Ratio (%) 0 20 40 60 80 100 # of Transmissions/Message 0 20 40 60 80 Avg. Delivery time(h) 0 50 100 150 Message TTL (d) Delivery 10 Relay Points (e) Transmissions 10 Relay Points (f) Delay 10 Relay Points Fig. 4. Performance comparison of, Relay-Bubble and in the NCCU dataset Indeed, the key for s cost-effectiveness is the combination of all those spatial and social features. The experiments show that Relay-Bubble overcomes in delivery ratio and delay, even with a few relay points. Since both algorithms forward messages based only on node s popularity, the usage of relay points, deployed at PoIs, enhances the chances of a message to get to the target community faster, but at the price of a higher network overhead. As described before, once the message gets to a relay point, works just like Relay-Bubble. Hence, the difference in the results is due to the usage of the node s radius of gyration (mobility awareness) in the forwarding policy, when the message is outside the destination community. Since nodes with a higher radius of gyration have a better geographical coverage, using it increases the chances of delivering the messages to a relay point with fewer re-transmissions, reducing the network overhead without compromising the delivery ratio and messages average delay. In the SWIM trace, s overhead becomes even lower than s. This is because in this scenario, with more nodes, and Relay-Bubble (both use popularity-based forwarding strategies when outside the destination community) tend to re-transmit more messages. On the other hand, s mobility-based forwarding maintains the network overhead lower when the network scale is increased. In contrast to the results presented in the NCCU trace, exhibits slower delivery in this scenario. It happens because the SWIM mobility model does not account for regions of higher popularity within the mobility trace. Indeed, the distribution of nodes positions in the SWIM trace is much more uniform than the one presented in real traces (see Figure 1(e)). This effect is confirmed when we look at Figures 5(c) and 5(f) and see that increasing the number of relay points from five to ten does not help to decrease the delay. It is worth mentioning that, even in this disadvantageous synthetic scenario with no PoIs, presents competitive results due to its low overhead. VI. FINAL REMARKS In this work, we have introduced the use of spatial information together with social awareness to improve the costeffectiveness of D2D opportunistic routing. We have discussed spatial and social properties of mobility and how these properties can be retrieved from mobility traces and used in opportunistic routing. As a proof of concept, we proposed SAM- PLER, a simple scheme that combines four different features: nodes popularity, individual mobility patterns, PoIs, and social communities. exemplifies how the combination of such features can provide significant improvements, enabling higher content delivery while reducing network overhead and average delivery time of messages.

IEEE COMMUNICATIONS MAGAZINE 8 Delivery Ratio (%) 0 20 40 60 80 100 # of Transmissions/Message 0 50 100 150 200 Avg. Delivery time(h) 0 50 100 150 Message TTL (a) Delivery 5 Relay Points (b) Transmissions 5 Relay Points (c) Delay 5 Relay Points Delivery Ratio (%) 0 20 40 60 80 100 # of Transmissions/Message 0 50 100 150 200 Avg. Delivery time(h) 0 50 100 150 Message TTL (d) Delivery 10 Relay Points (e) Transmissions 10 Relay Points (f) Delay 10 Relay Points Fig. 5. Performance comparison of, Relay-Bubble and in the SWIM synthetic trace Considering the combination of spatial and social domains, several limitations and open issues arise. We here highlight the importance of collecting and publishing new large-scale real-world mobility traces, what is not the current scenario. This would allow further investigation of combined socialaware and spatial-aware strategies and better validation of such strategies under different scenarios. Considering social features, we need to further investigate new features with advantages to communities. In [10], for instance, the use of group meetings awareness is proposed as an alternative measure of social context, because community detection is a hard task to perform in real-world distributed networks. Another future direction is to explore PoIs semantics, i.e., motivations behind users transitions among PoIs. Finally, as in any system that relies on geo-positioning, strategies must be proposed to obtain the position information in an energyefficient way. ACKNOWLEDGMENT The authors would like to thank CNPq and FAPEMIG agencies for their financial support. REFERENCES [1] Y. Li, T. Wu, P. Hui, D. Jin, and S. Chen, Social-aware D2D Communications: Qualitative Insights and Quantitative Analysis, IEEE Communications Magazine, vol. 52, no. 6, pp. 150 158, 2014. [2] I. O. Nunes, P. Vaz de Melo, and A. A. F. Loureiro, Leveraging D2D Multi-Hop Communication Through Social Group Meetings Awareness, IEEE Wireless Communications Magazine, vol. 23, no. 4, pp. 12 19, 2016. [3] P. Hui, J. Crowcroft, and E. Yoneki, : Social-based Forwarding in Delay-Tolerant Networks, IEEE Transactions on Mobile Computing, vol. 10, no. 11, pp. 1576 1589, 2011. [4] T.-C. Tsai and H.-H. Chan, NCCU Trace: Social-network-aware Mobility Trace, IEEE Communications Magazine, vol. 53, no. 10, pp. 144 149, 2015. [5] S. Kosta, A. Mei, and J. Stefa, Large-Scale Synthetic Social Mobile Networks with SWIM, IEEE Transactions on Mobile Computing, vol. 13, no. 1, pp. 116 129, 2014. [6] T. Spyropoulos, K. Psounis, and C. S. Raghavendra, Spray and Wait: An Efficient Routing Scheme for Intermittently Connected Mobile Networks, in Proceedings of the 2005 ACM SIGCOMM Workshop on Delay-tolerant Networking. ACM, 2005, pp. 252 259. [7] L. Amaral, R. Sofia, P. Mendes, and W. Moreira, Oi!-Opportunistic Data Transmission Based on Wi-Fi Direct, in 2016 IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS). IEEE, 2016, pp. 578 579. [8] W. Moreira, P. Mendes, and S. Sargento, Social-aware Opportunistic Routing Protocol Based on Users Interactions and Interests, in International Conference on Ad Hoc Networks. Springer, 2013, pp. 100 115. [9] A. Lindgren, A. Doria, and O. Schelén, Probabilistic Routing in Intermittently Connected Networks, ACM SIGMOBILE Mobile Computing and Communications Review, vol. 7, no. 3, pp. 19 20, 2003. [10] I. O. Nunes, C. Celes, P. Vaz de Melo, and A. A. F. Loureiro,

IEEE COMMUNICATIONS MAGAZINE [11] [12] [13] [14] [15] GROUPS-NET: Group meetings aware routing in multi-hop D2D networks, Computer Networks, vol. 127, pp. 94-108, 2017. V. F. Mota, F. D. Cunha, D. F. Macedo, J. M. Nogueira, and A. A. Loureiro, Protocols, Mobility Models and Tools in Opportunistic Networks: A Survey, Computer Communications, vol. 48, pp. 5 19, 2014. G. Palla, I. Dere nyi, I. Farkas, and T. Vicsek, Uncovering the Overlapping Community Structure of Complex Networks in Nature and Society, Nature, vol. 435, no. 7043, pp. 814 818, 2005. L. Peel, Estimating Network Parameters for Selecting Community Detection Algorithms, in 2010 13th Conference on Information Fusion (FUSION). IEEE, 2010, pp. 1 8. M. C. Gonzalez, C. A. Hidalgo, and A.-L. Barabasi, Understanding Individual Human Mobility Patterns, Nature, vol. 453, no. 7196, pp. 779 782, 2008. S. Shahbazi, S. Karunasekera, and A. Harwood, Improving Performance in Delay/Disruption Tolerant Networks Through Passive Relay Points, Wireless Networks, vol. 18, no. 1, pp. 9 31, 2012. Ivan Oliveira Nunes is currently a networked systems Ph.D. student at the University of California Irvine (UCI). He received his M.Sc. degree in Computer Science from the Federal University of Minas Gerais (UFMG), Brazil, in 2016, and his B.Sc. degree in Computer Engineering from the Federal University of Espirito Santo (UFES), Brazil, in 2014. His current research interests include networking, mobile and ubiquitous computing, security, and embedded systems. Clayson Celes is currently a computer science Ph.D. candidate at Federal University of Minas Gerais (UFMG), Brazil. He received his M.Sc. degree in Computer Science from UFMG in 2013 and his B.Sc. degree in Computer Science from the State University of Ceara (UECE), Brazil, in 2010. His research areas are mobile computing, vehicular networks, and ubiquitous computing. Igor Nunes is currently a Computer Engineering student at Federal University of Espirito Santo (UFES), Brazil. His research interests include statistical models, machine learning, computing systems and ubiquitous computing. Pedro O.S. Vaz de Melo is an assistant professor in the Computer Science Department (DCC) of Federal University of Minas Gerais (UFMG). He got his Ph.D. from UFMG with a one year period as a visiting researcher in Carnegie Mellon University and a five-month period as a visiting researcher at INRIA Lyon. His research interest is mostly focused on knowledge discovery and data mining in complex and distributed systems. View publication stats 9 Antonio A. F. Loureiro holds a Ph.D. degree in Computer Science from the University of British Columbia, Canada. Currently, he is a full professor at UFMG. His main research areas are mobile computing, vehicular networks, wireless sensor networks, and distributed algorithms. In the last 15 years he has published regularly in international conferences and journals related to those areas, and also presented keynotes and tutorials at international conferences.